Resolved alibi bias issue by reverting to PA v1. #438
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Requires associated changes on vllm-fork PR
This reverts forward_decode to use paged_attention_v1 for models which required ALiBI. This was broken in PR #169.
Below you can see the results on mpt-7B model which uses ALiBI. The first example is the current code on main. Second is an alternate fix which adds ALiBI to flat_pa. And the third is this PR which reverts the decode attention back to paged_attention_v1 before #169 which produces the best results in my opinion.